TDM28 Model Design DRAFT Draft Version: September 24, 2026

Document Purpose

● Living document representing current thinking on TDM28 Design.

○ Final design to be settled with Model Support consultant

● Unspecified details are assumed to be consistent with TDM23.

● Highlights unresolved questions and captures key decisions.

○ Unresolved questions highlighted in blue

● Does not assign responsible parties (CTPS, Consultant, or Other Agencies).

● Serves as the starting point for the TDM28 Structures and Performance Report (e.g., https://ctps.org/pub/tdm23_sc/tdm23.2.1/TDM23.2.1_Structures%20and%20Performanc e.pdf).

Section 1: Background and Constraints

Travel Modeling at CTPS CTPS maintains a statewide travel demand model to support the Boston Region MPO, MassDOT, and MPOs and RPAs across Massachusetts.

Modeling practice prioritizes utility: models are only as complex as necessary, accessible to stakeholders, and easy to run, understand, debug, and investigate for predictive and exploratory uses.

The current travel demand model, TDM23, re-estimated all parameters, refactored network structures, produced new scripts, and leveraged disaggregate socio-economic projections.

Python and JavaScript tools support visualization, post-processing, and calibration. CTPS evaluated TMIP-EMAT capabilities for transportation planning through experimental tests.

Made publicly available in early 2024, TDM23 has been accessed by over 30 organizations.

Updated releases incorporated input refinements, bug fixes, and a Utility Platform UI for pre/post-processors and visualizations. A TDM23 3.0 release is underway to recalibrate the base year using recent Massachusetts household travel survey data.

Predictive vs. Exploratory

CTPS is transitioning to a dual-platform Travel Demand Modeling Program, splitting activities into statutory (predictive) and exploratory paradigms. This fulfills statutory mandates with defensible predictive tools while equipping planners with flexible disaggregate models to evaluate uncertainties.

https://ctps.org/pub/tdm23_sc/tdm23.2.1/TDM23.2.1_Structures%20and%20Performance.pdf https://ctps.org/pub/tdm23_sc/tdm23.2.1/TDM23.2.1_Structures%20and%20Performance.pdf

The statutory platform (TDM28) satisfies legal, regulatory, and certification requirements for the MPO, including regional air quality conformity determinations, MEPA/EIS environmental reviews, federal grant applications, and 2028 Long-Range Transportation Plan (LRTP) projections.

The exploratory platform (TDMx) supports decision-making under deep uncertainty (e.g., AV adoption, dynamic road pricing, hybrid telework shifts, e-commerce, climate resilience). Rather than point-estimate forecasts, TDMx evaluates scenario ranges to assess system responses to policy and market shifts.

Statutory Model Constraints The Statutory Model prioritizes model stability, strict versioning, auditability, and computational throughput to deliver reproducible, single-estimate projections across project alternatives.

Schedule TDM28 supports the 2028 Boston Region MPO LRTP and statewide LRTPs. Model design, development, and calibration must conclude by October 2027 to enable scenario evaluations.

Anticipated schedule and major milestones:

A Design completed

B Models estimated

C Base year inputs developed

D Calibrated

E Horizon year inputs developed

F Full Documentation / User Guide

G Public share

TDM28 Design Draft - 20260924

Available and Easy to Use and Maintain

● Documented and versioned following TDM23 standards.

● Simplified population inputs to streamline alternative scenario definitions.

● Open-source code and data structures where viable.

● Calibrated thoroughly against available observations.

● Fast execution, low memory footprint, and storage-efficient scenario management.

To ensure operational viability across RPAs, partner MPOs, and consultants, TDM28 maintains an aggregate, trip-based structure in TransCAD and Python, optimized to execute in under 8 hours on standard workstations.

Section 2: Structures

Overview

● Networks and Land Use: 2026 Base Year

● Costs: 2026 Year Dollars

● Platform Implementation: Aggregate, trip-based model structure in TransCAD and

Python, targeted to run under 8 hours on standard agency workstations.

Geographic Extent

● Massachusetts: TAZ boundaries re-engineered to align with 2020 Census Blocks.

● New Hampshire: Revised extent to align with NHC RPC model boundaries using a simplified TAZ structure.

● Rhode Island: Simplified TAZ structure.

● Connecticut (Unresolved): Evaluating potential extent expansion into Connecticut (south of Springfield, MA).

Temporal Representation Temporal resolution expanded to 5 time periods by adding an Evening (EV) period:

Time Periods:

● AM Peak: 6:00 AM - 9:00 AM

● Midday (MD): 9:00 AM - 2:00 PM

● PM Peak: 2:00 PM - 6:00 PM

● Evening (EV): 6:00 PM - 9:00 PM

● Night (NT): 9:00 PM - 6:00 AM

Source: CTPS preliminary analysis of 2024 HH Survey compared to 2011 HH Survey

Inputs TDM23 initiated trip generation at the Census Block level before aggregating to 5,849 TAZs.

TDM28 will operate directly at the TAZ level from inception, utilizing re-engineered TAZ boundaries aligned with 2020 Census Blocks and rationalized external zone extents in NH and

RI.

To prevent aggregation bias in non-linear models without relying on fully disaggregate synthetic populations, TDM28 will use market segmentation. Synthetic population data from UrbanSim and PopulationSim will be aggregated into zonal cross-classifications across two primary dimensions:

● Household Income: Standard income segments consistentwith TDM23.

Household Composition: Lifecycle groups capturing variation in driving-age adults (working and non-working) and children.

Note: Models may still be estimated at the disaggregate level provided explanatory variables are restricted to zonal or cross-segmented population metrics.

SocioEconomic Data Population

● Massachusetts: 2026 estimates from UrbanSim (maintained by MAPC).

● Outside Massachusetts: Disaggregate populations generated via PopulationSim using out-of-state controls.

○ Potential Aggregation Segments: Households aggregated by Size, Income, Presence of Children, and Workers.

Employment

● Evaluate potential aggregation of 10 Super Sectors based on correlation testing and predictive value.

Transit Access Density (Area Type)

● Maintain existing TDM23 definitions.

Walk / Bike Conditions

● Unresolved: Evaluating replacing intersection density with crosswalk density derived from UMass Lowell identification methods.

Parking Costs

● Update parking cost assumptions from 2024 to 2026 values.

Direct Trip Inputs

External Stations

● Identify new external stations matching modified model extents.

● Update AWDT and auto/truck split ratios.

● Update External-External (EE) seed tables.

● Unresolved: Identify additional stations based on RPA planning needs.

Airport Ground Access

● Update enplanement counts and purpose splits.

● Unresolved: Incorporate adjacent regional airports (Manchester, TF Green, Hartford).

Networks

Roadway

● Maintain standard network attributes.

● Update centroid connectors for new TAZ geometry.

● Integrate latest regional networks from RI, NH, and CT.

Transit

● Maintain standard transit network attributes.

● Update to 2026 service levels and calculate route headways across the 5 revised time periods.

Non-Motorized

● Unresolved: Review non-motorized network skim differences to confirm operational utility.

Modal Characteristics

TNC fares in TDM23 were estimated from sampled trip requests, while wait times served as a calibration parameter.Unresolved: revise/resstructure TNC wait times and fares leveraging updated DPU data.

Model Components TDM28 maintains a trip-based architecture similar to TDM23, refined to enhance overall model usability.

Core Demand (Resident Passenger)

Vehicle Availability

● Retain current alternatives (0 vehicles, vehicles < driving-age adults, vehicles >= driving-age adults); re-estimate using aggregated demographic inputs.

● Current vehicle availability alternatives correlate strongly with distinct mode share patterns.

Source: CTPS preliminary analysis of 2024 HH Survey compared to 2011 HH Survey

Explanatory variables consist of zonal household income, household composition, parking costs, and transit service quality metrics.

An alternative specification predicts discrete vehicle counts (0, 1, 2, 3, 4+ autos) before assigning auto sufficiency categories against driving-age adults (enabling direct ACS validation), though retaining the current auto sufficiency formulation is preferred to minimize unnecessary structural changes.

Trip Generation To prevent aggregation bias, trip generation models predict daily household trip counts rather than zonal averages using negative binomial regression. While unweighted estimation minimizes standard errors, previous TDM23.3 re-estimation demonstrated that applying linked trip weights as household weights is necessary to match empirical calibration targets and prevent trip underestimation.

Trip Distribution Destination choice retains mode choice logsums as accessibility measures (with logsum coefficients constrained to <= 1.0). Auto-sufficiency segmentation is retained for downstream destination and mode choice models (plus income segmentation for Home-Based Work).

For Destination Choice models:

● Implement district-to-district constants across 10–20 key regional districts to calibrate macro-level flows (e.g., river crossings and major highway corridors) using Replica data.

● District-to-district constants capture systematic spatial unobservables during estimation without functioning as arbitrary k-factors, ensuring mode choice logsum accessibility variables maintain statistical significance.

Mode Choice

● Mode choice structures remain substantially unchanged under aggregate market segmentation, with primary modifications concentrated on conditional non-home-based trip handling.

Source: CTPS preliminary analysis of 2024 HH Survey compared to 2011 HH Survey

Non-Home Based Trips TDM28 addresses traditional trip-based limitations by executing non-home-based (NHB) generation, distribution, and mode choice sequentially based on home-based (HB) trip outcomes rather than in parallel.

Sequential NHB implementation enhances sensitivity for:

- Connecting HB and NHB trip modes (transit HB trips correlate with non-auto NHB trips;

SOV HB trips correlate with auto NHB trips).

- Evaluating transit service adjustments and Transit-Oriented Development (TOD) shifts

(e.g., increased HBW transit trips to urban hubs directly reduce local NHB auto trips).

- Modeling cordon pricing and road pricing impacts on internal NHB trips within pricing zones.

- Capturing regional development impacts by linking NHB trips directly to changes in site-specific primary attractions.

Unresolved: Potential Implementation Workflow (applicable to Work, Social/Rec, Personal Business; simple parallel structures retained for School, University, and Special Generators):

1. Generate HB trips using market segmentation and collapse production zones down to auto sufficiency (and income for HBW).

2. Split HB trips by peak/off-peak periods and execute HB mode choice and destination choice (where HB attraction zones act as destination ends).

3. Aggregate trips by HB purpose, attraction TAZ, primary mode, and peak/off-peak period; drop auto sufficiency and income dimensions.

4. Directionally factor HB trips into "from-home" and "to-home" vectors using travel survey factors combined with time-of-day splits.

5. Estimate NHBW and NHBNW trip generation from HB attraction zones (acting as NHB production zones) using primary trip purpose, arrival mode, time of day, and directionality as explanatory variables (excluding local residential demographics).

Use trip generation models to generate NHBW and NHBNW trips from each HB attraction zone—which are the production zones for NHB trips. The explanatory variables can include the purpose of the trip to the zone (e.g. HBW work trips generate the most NHBW trips), the mode used to the zone (e.g. auto tours tend to have more trips per tour than non-auto tours), the peak versus off-peak period, and the from-home versus to-home direction. Population characteristics of the HB attraction zone are not included because the people making the associated NHB trips generally do not live in that zone.

6. Estimate and apply NHB destination choice and mode choice models segmented by incoming primary mode, time period, and directionality.

7. Aggregate all HB and NHB trip matrices by O-D zone pair, mode, and time period for assignment.

Auxiliary Demand

HBU and On-Campus College Students

University Students (HBU & On-Campus): Incorporating explicit university student travel models to address group quarters omissions in TDM23.

Truck trips Commercial Trucks: Refining commercial vehicle models to improve light, medium, and heavy-duty truck representation.

Assignment and Feedback Unresolved:To eliminate numerical noise and speed up highway equilibrium assignment convergence, TDM28 will implement a matrix integerization (rounding) routine for Drive Alone (DA) and Shared Ride (SR) trip tables prior to assignment:

1. Round all row and column matrix totals.

2. Round all individual cell values down to integers and store fractional residuals.

3. Calculate row and column residual targets by subtracting floored totals from rounded targets.

4. Rank order cells by residual magnitude.

5. Increment highest-residual cells by +1 and adjust row/column residual targets until all row and column constraints close.

Feedback

● Maintain highway AM and Midday (MD) feedback assignment and skimming.

● Maintain transit Park-and-Ride (PnR) AM feedback assignment and skimming.

Validation TDM23 demonstrated systematic errors, including highway over-prediction in dense urban cores, under-prediction in outer ring suburbs, commuter rail transit biases, and evening TNC distortions. TDM28 focuses on parameter re-estimation using recent household survey data (2024–2025) and comprehensive recalibration against post-pandemic (2024–2026) counts and boardings.

Primary data sources for validation include:

● 2024–2025 Household Travel Survey.

● MBTA Ridership data

● Replica OD flows and AWDT estimates

● Department of Public Utilities (DPU) TNC data.

Implementation TDM28 leverages TDM23 code structures in TransCAD and Python, refactored to conform to contemporary software best practices. Refactor TDM23 so that model execution originates in TransCAD, Python code is safe in both embedded and subprocess runtimes, and model users can access data directly through familiar files rather than a database tool.

Specific Implementation Tasks:

● Maintain folder hierarchy and naming conventions consistent with TDM23.

● Upgrade to TransCAD 11 (if available) and restructure Python calls to execute natively.

● Refactor Python code for speed and maintainability (e.g. replace .apply() with vectorized pandas operations).

● Execute Python components natively within TransCAD 10/11.

● Replace embedded SQLite databases with accessible, standardized flat-file formats

(e.g., CSV) to facilitate external querying and script-based post-processing

● [more details in separate refactoring doc]